Compute Unified Architecture (CUDA) has been a revolutionary technology, enabling general-purpose processing and tasks to run on NVIDIA GPUs. Its widespread adoption, especially for machine learning and generative AI workloads, has solidified NVIDIA GPUs as the go-to choice for developers and researchers. However, this dominance has left AMD GPUs lagging behind, lacking native support for CUDA applications. Now, a new startup, Spectral Compute, has introduced a game-changer: the SCALE toolkit. SCALE empowers developers to compile CUDA applications for AMD’s RDNA GPUs without any modifications to the CUDA program or its build system. This breakthrough enables a “write once, run anywhere” paradigm, long established for CPUs, to finally extend to GPUs. According to Spectral Compute CEO Michael Sondergaard, their mission is to bridge the compatibility gap between CUDA and other hardware vendors. SCALE achieves this by effectively impersonating the NVIDIA CUDA Toolkit, ensuring compatibility with existing tools and scripts. The toolkit has been in development for seven years and does not rely on NVIDIA’s code, making it a truly independent solution. This independence could be crucial in avoiding potential legal complications. Spectral Compute has successfully tested SCALE with various applications, including Blender, Llama-cpp, XGboost, FAISS, and GOMC, all running seamlessly on AMD GPUs. This development could mark a significant shift in the GPU landscape, opening up new possibilities for AMD GPUs and giving developers greater flexibility in their hardware choices.
AMD GPUs Gain CUDA Compatibility with SCALE Toolkit
